pith. machine review for the scientific record. sign in

arxiv: 2605.07680 · v1 · submitted 2026-05-08 · 🌌 astro-ph.HE

Recognition: 2 theorem links

· Lean Theorem

Advance warning of γ-ray blazar flares from textit{Fermi}-LAT light curves: a strictly causal machine-learning backtest

Sikandar Akbar, Zahir Shah

Authors on Pith no claims yet

Pith reviewed 2026-05-11 02:08 UTC · model grok-4.3

classification 🌌 astro-ph.HE
keywords blazar flaresFermi-LAT light curvesmachine learning predictiongamma-ray astronomycausal modelinglogistic regressionflare forecastingvariability features
0
0 comments X

The pith

A polynomial logistic regression model on Fermi-LAT light curves issues gamma-ray blazar flare alerts up to 4.5 days before onset.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether long-term Fermi satellite monitoring of blazars can reveal patterns that precede bright gamma-ray flares. It divides each light curve into 365-day trailing windows, extracts 42 variability features from each, and trains classifiers to predict whether flare activity will occur within the next 90 days or a new flare onset within the next 45 days. All scaling, threshold choice, and validation stay strictly before a fixed cutoff date to prevent any use of future information. When the best model is applied to one target source after training on 13 other bright blazars, it reaches ROC AUC 0.891 on held-out data, flags most pre-flare windows, and produces alerts before both tested flare episodes. If the patterns generalize, observers could prepare multi-wavelength follow-up before the flare becomes obvious in the gamma-ray band.

Core claim

Variability features measured in 365-day trailing windows from 3-day binned Fermi-LAT light curves contain predictive information about upcoming flares. A polynomial logistic regression classifier, trained on 13 auxiliary blazars and tested on 4FGL J1048.4+7143 with all calibration performed only on pre-MJD-60000 data, achieves ROC AUC 0.891 and average precision 0.396 for 90-day WATCH predictions, recovers 18 of 21 positive windows, and issues final alerts 4.5 and 2.5 days before the two held-out flare onsets, while the corresponding WATCH-active periods begin 88.5 and 72.5 days earlier.

What carries the argument

The strictly causal pipeline that samples 365-day trailing windows, computes 42 variability features per window, and trains separate WATCH (90-day flare activity) and TRIGGER (45-day flare onset) classifiers on auxiliary blazars while restricting all scaling and validation to pre-cutoff data.

If this is right

  • Long-term Fermi light curves contain usable signals about the build-up to blazar flares.
  • Polynomial logistic regression yields the strongest held-out ranking performance among the classifiers tested.
  • The WATCH model recovers 86 percent of pre-flare windows at the chosen threshold.
  • Alerts appear days before both held-out flare onsets, with active periods beginning more than two months earlier.
  • The same framework produces weaker but still positive ranking for the shorter TRIGGER horizon.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Feature-importance analysis on the trained model could highlight which variability measures best signal impending flares.
  • If the learned patterns are not source-specific, the same coefficients might forecast flares in additional blazars without retraining.
  • Combining the WATCH state with simultaneous lower-energy monitoring could improve the reliability of advance alerts for coordinated campaigns.
  • The causal-window approach supplies a reusable template for forecasting other transient events that have dense long-term monitoring.

Load-bearing premise

The 42 variability features from 365-day windows genuinely reflect physical processes that precede flares, and the model trained on 13 auxiliary blazars generalizes to the target source without capturing source-specific noise.

What would settle it

A new flare episode in the same source where the model never enters the WATCH state or fails to issue an alert before onset, or markedly lower performance when the identical pipeline is run on additional independent blazars.

Figures

Figures reproduced from arXiv: 2605.07680 by Sikandar Akbar, Zahir Shah.

Figure 1
Figure 1. Figure 1: Schematic of the strictly causal blazar flare￾forecasting pipeline. The light curves first pass through qual￾ity cuts, and the TRAIN-only quiescent level Fq is defined for later normalization of flux-dependent features. Bayesian Blocks are then computed in two ways: TRAIN-BB uses only data up to tcut = MJD 60000 for training labels, while FULL-BB is used only to score the held-out target stream. Rolling 36… view at source ↗
Figure 2
Figure 2. Figure 2: Bayesian-Blocks diagnostics for the target source. In each panel, the upper plot shows the 3-d Fermi-LAT light curve together with the Bayesian-Blocks representation. Each semi-transparent shaded strip spans one Bayesian block in time and extends from zero flux to the block-mean flux: red/pink strips mark blocks classified as flaring, whereas light-blue strips mark non-flaring blocks. The red dashed line i… view at source ↗
Figure 3
Figure 3. Figure 3: WATCH-only soft-voting ensemble for the target source. For each rolling window, the raw calibrated WATCH probabilities from logistic regression, polynomial logistic re￾gression, and random forest are averaged to form the ensem￾ble WATCH probability. The top panel shows the 3-d binned Fermi-LAT flux; the salmon-shaded bands mark the FULL￾BB flare intervals, and the navy dashed vertical line marks the train/… view at source ↗
Figure 4
Figure 4. Figure 4: Held-out WATCH diagnostic curves for the PLR model. In the ROC panel, the blue step-like curve is obtained by varying the WATCH probability threshold across the held-out windows and plotting the true-positive rate against the false￾positive rate at each step. The black dashed diagonal marks the random-ranking case, and the legend gives the area under the curve (AUC = 0.891). In the precision–recall panel, … view at source ↗
Figure 5
Figure 5. Figure 5: PLR alert-state timeline. The top panel shows the 3-d binned Fermi-LAT flux as black points; salmon shad￾ing marks FULL-BB flare intervals, and the navy dashed vertical line marks the train/test boundary at Tboundary = MJD 59910. The lower panel shows the causal seven-window trailing means of the WATCH score (solid blue) and the TRIGGER score (dark-red dashed). Filled green circles mark windows that are WA… view at source ↗
read the original abstract

Long-term \textit{Fermi}-LAT monitoring makes it possible to ask whether a blazar light curve shows signs of an upcoming flare before the flare becomes obvious in the $\gamma$-ray emission. We present a strictly causal machine-learning framework for forecasting $\gamma$-ray blazar flares from 3-d binned LAT light curves. Flare intervals are identified with Bayesian Blocks, and each light curve is sampled with 365-d trailing windows from which 42 variability features are measured. We train separate WATCH and TRIGGER models: WATCH predicts whether flare activity will appear within the next 90 d, while TRIGGER predicts whether a new flare onset will occur within the next 45 d. To avoid temporal leakage, all scaling, calibration, threshold selection, and validation use only the pre-cutoff data before MJD 60000. We apply the method to the FSRQ 4FGL\,J1048.4$+$7143, using 13 bright blazars as auxiliary training sources. Among logistic regression, polynomial logistic regression, and random forest classifiers, polynomial logistic regression gives the strongest held-out WATCH performance, with ROC AUC $=0.891$, average precision $=0.396$, and a block-permutation probability $p_{\rm perm}=0.006$. At the selected WATCH threshold, it recovers 18 of the 21 positive windows in the held-out WATCH set, corresponding to a recall of 0.86. The same model also gives the best held-out TRIGGER ranking, with TRIGGER AUC $=0.770$ and TRIGGER AP $=0.123$, although no reliable pre-onset TRIGGER alert is obtained. The WATCH state appears before both held-out flare episodes, with final alerts 4.5 and 2.5 d before onset. The corresponding broader WATCH-active periods begin 88.5 and 72.5 d before flare onset. These results suggest that long-term {\fermi} light curves contain useful predictive information about the build-up to blazar flares.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper presents a strictly causal machine-learning framework to forecast γ-ray blazar flares from 3-day binned Fermi-LAT light curves. Flare intervals are defined via Bayesian Blocks; 42 variability features are extracted from 365-day trailing windows. Separate WATCH (90-day flare-activity horizon) and TRIGGER (45-day onset horizon) classifiers are trained on 13 auxiliary blazars with all scaling, calibration, and threshold selection performed exclusively on pre-MJD 60000 data; the models are then evaluated on the post-cutoff light curve of 4FGL J1048.4+7143. Polynomial logistic regression yields the best held-out WATCH performance (ROC AUC = 0.891, recall = 0.86 on 21 positive windows, block-permutation p = 0.006), issuing final alerts 4.5 d and 2.5 d before the two observed flares.

Significance. If the reported generalization holds, the work would supply a practical, low-latency tool for scheduling multi-wavelength observations of blazar flares. The strictly causal pipeline, pre-cutoff validation, and permutation test are methodological strengths. However, the modest average precision (0.396), the restriction to a single target source, and the absence of cross-source ablation tests limit the immediate astrophysical impact; broader validation would be required before the method could be considered a robust predictor of flare build-up.

major comments (3)
  1. [Results (held-out WATCH evaluation)] The central claim that the 42 variability features encode transferable physical information about flare precursors rests on training on 13 auxiliary blazars and testing on only one target (4FGL J1048.4+7143) after MJD 60000. No leave-one-source-out cross-validation, feature-ablation study, or source-wise permutation test is reported to demonstrate that performance is not driven by source-specific variability statistics or selection biases in the auxiliary sample.
  2. [Results (WATCH performance metrics)] The 21 positive WATCH windows are clustered around only two flare episodes; combined with the low average precision of 0.396, this raises the possibility that the reported AUC = 0.891 and recall = 0.86 reflect limited-sample idiosyncrasies rather than robust precursors. A sensitivity analysis on window length or feature subset would be needed to substantiate the claim.
  3. [Results (TRIGGER model)] The TRIGGER model is substantially weaker (AUC = 0.770, AP = 0.123) and yields no reliable pre-onset alerts, yet the manuscript’s headline result and discussion focus on the WATCH model. The discrepancy between the two tasks should be quantified and discussed as it directly affects the practical utility for flare-onset forecasting.
minor comments (2)
  1. [Methods (feature extraction)] Clarify in the methods whether the 42 features include any explicit flux or spectral indices that could inadvertently leak information across the MJD 60000 boundary despite the causal windowing.
  2. [Abstract] The abstract states that the WATCH state appears before both held-out flares; add a brief statement on the false-positive rate during the long WATCH-active periods (88.5 d and 72.5 d) to give a complete picture of operational cost.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the strengths of our strictly causal pipeline and pre-cutoff validation. We address each major comment below. Where the comments identify genuine limitations or opportunities for clarification, we have revised the manuscript accordingly.

read point-by-point responses

  1. Referee: The central claim that the 42 variability features encode transferable physical information about flare precursors rests on training on 13 auxiliary blazars and testing on only one target (4FGL J1048.4+7143) after MJD 60000. No leave-one-source-out cross-validation, feature-ablation study, or source-wise permutation test is reported to demonstrate that performance is not driven by source-specific variability statistics or selection biases in the auxiliary sample.

    Authors: We selected the single held-out target because it is the only source with sufficient post-MJD 60000 coverage containing multiple well-defined flares, allowing a true temporal hold-out. Leave-one-source-out cross-validation on the auxiliary set would test intra-sample generalization rather than the intended out-of-distribution transfer to a new source, which is the core scientific claim. We have added a source-wise permutation test (randomly reassigning auxiliary labels while preserving temporal structure) to the revised results section; this yields p = 0.008, supporting that performance is not driven by any single auxiliary source. A full feature-ablation study was not performed due to computational cost, but we now report the top-10 features ranked by logistic-regression coefficients and note the absence of a complete ablation as a limitation in the discussion. revision: partial

  2. Referee: The 21 positive WATCH windows are clustered around only two flare episodes; combined with the low average precision of 0.396, this raises the possibility that the reported AUC = 0.891 and recall = 0.86 reflect limited-sample idiosyncrasies rather than robust precursors. A sensitivity analysis on window length or feature subset would be needed to substantiate the claim.

    Authors: The clustering around two flares is an unavoidable consequence of the post-cutoff data for this source; we have added explicit language in the results and discussion stating that the held-out sample contains only two flare episodes and that the reported metrics should be viewed as preliminary. We performed an internal sensitivity check on window lengths (300–400 d) during model development and found AUC stable within ±0.03; we now include a short sensitivity table in the supplementary material showing that the top-ranked features remain predictive when the lowest-importance 20 % of features are removed. We agree that broader validation on sources with more flare events is required before claiming robustness. revision: partial

  3. Referee: The TRIGGER model is substantially weaker (AUC = 0.770, AP = 0.123) and yields no reliable pre-onset alerts, yet the manuscript’s headline result and discussion focus on the WATCH model. The discrepancy between the two tasks should be quantified and discussed as it directly affects the practical utility for flare-onset forecasting.

    Authors: We agree that the performance gap between WATCH and TRIGGER requires more explicit treatment. In the revised discussion we now quantify the difference (ΔAUC = 0.121, ΔAP = 0.273) and explain that WATCH identifies extended flare-active periods (useful for scheduling), while TRIGGER attempts precise 45-day onset prediction on a much sparser positive class. We have added a paragraph discussing why the shorter horizon and lower event density make TRIGGER inherently harder, and we temper the practical-utility claims accordingly, noting that reliable onset alerts are not yet achieved. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML prediction on time series with held-out evaluation

full rationale

The paper describes a supervised machine-learning pipeline that extracts 42 variability features from 365-day trailing windows of Fermi-LAT light curves, trains classifiers (logistic regression, polynomial logistic regression, random forest) on 13 auxiliary blazars, and evaluates strictly on held-out post-MJD-60000 data for one target source. Flare labels are obtained via Bayesian Blocks, and all scaling, calibration, and threshold choices are confined to pre-cutoff data to enforce causality. No equations, derivations, or self-citations are invoked to define or justify the target quantities; the reported AUC, recall, and alert lead times are direct empirical outcomes of the trained models on independent test windows. The pipeline therefore contains no self-definitional, fitted-input-renamed-as-prediction, or self-citation-load-bearing steps.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard domain assumptions about light-curve variability and flare identification plus a small number of hyperparameters chosen from pre-cutoff data.

free parameters (3)
  • 365-day trailing window length
    Fixed length chosen for sampling each light curve segment
  • 90-day and 45-day prediction horizons
    Defined for the WATCH and TRIGGER tasks
  • WATCH decision threshold
    Selected on pre-cutoff data to achieve target recall
axioms (2)
  • domain assumption Bayesian Blocks correctly partitions the light curve into flare and quiescent intervals
    Used to label positive and negative training windows
  • domain assumption The 42 variability features extracted from each window are sufficient to capture flare-precursor information
    Central to the feature-based classification approach

pith-pipeline@v0.9.0 · 5683 in / 1518 out tokens · 57003 ms · 2026-05-11T02:08:51.430560+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Reference graph

Works this paper leans on

61 extracted references · 61 canonical work pages · 1 internal anchor

  1. [1]

    A., Ackermann, M., Ajello, M., et al

    Abdo, A. A., Ackermann, M., Ajello, M., et al. 2010, ApJ, 722, 520, doi: 10.1088/0004-637X/722/1/520

    work page doi:10.1088/0004-637x/722/1/520 2010

  2. [2]

    , keywords =

    Abdollahi, S., Ajello, M., Baldini, L., et al. 2023, ApJS, 265, 31, doi: 10.3847/1538-4365/acbb6a

    work page doi:10.3847/1538-4365/acbb6a 2023

  3. [3]

    2011, ApJ, 743, 171, doi: 10.1088/0004-637X/743/2/171

    Ackermann, M., Ajello, M., Allafort, A., et al. 2011, ApJ, 743, 171, doi: 10.1088/0004-637X/743/2/171

    work page doi:10.1088/0004-637x/743/2/171 2011

  4. [4]

    2015, ApJL, 813, L41, doi: 10.1088/2041-8205/813/2/L41

    Ackermann, M., Ajello, M., Albert, A., et al. 2015, ApJL, 813, L41, doi: 10.1088/2041-8205/813/2/L41

    work page doi:10.1088/2041-8205/813/2/l41 2015

  5. [5]

    G., Marscher, A

    Agudo, I., Jorstad, S. G., Marscher, A. P., et al. 2011, ApJL, 726, L13, doi: 10.1088/2041-8205/726/1/L13

    work page doi:10.1088/2041-8205/726/1/l13 2011

  6. [6]

    G., Bazer-Bachi, A

    Aharonian, F., Akhperjanian, A. G., Bazer-Bachi, A. R., et al. 2007, ApJL, 664, L71, doi: 10.1086/520635

    work page doi:10.1086/520635 2007

  7. [7]

    2026, Journal of High Energy Astrophysics, 53, 100608, doi: https://doi.org/10.1016/j.jheap.2026.100608

    Akbar, S. 2026, Journal of High Energy Astrophysics, 53, 100608, doi: https://doi.org/10.1016/j.jheap.2026.100608

    work page doi:10.1016/j.jheap.2026.100608 2026

  8. [8]

    2025, PhRvD, 112, 063061, doi: 10.1103/zxgv-fzv5

    Akbar, S., Shah, Z., Misra, R., Boked, S., & Iqbal, N. 2025, PhRvD, 112, 063061, doi: 10.1103/zxgv-fzv5

    work page doi:10.1103/zxgv-fzv5 2025

  9. [9]

    2024, ApJ, 977, 111, doi: 10.3847/1538-4357/ad8ddb

    Akbar, S., Shah, Z., Misra, R., & Iqbal, N. 2024, ApJ, 977, 111, doi: 10.3847/1538-4357/ad8ddb

    work page doi:10.3847/1538-4357/ad8ddb 2024

  10. [10]

    2007, ApJ, 669, 862, doi: 10.1086/521382 Astropy Collaboration, Price-Whelan, A

    Albert, J., Aliu, E., Anderhub, H., et al. 2007, ApJ, 669, 862, doi: 10.1086/521382 Astropy Collaboration, Price-Whelan, A. M., Lim, P. L., et al. 2022, ApJ, 935, 167, doi: 10.3847/1538-4357/ac7c74

    work page doi:10.1086/521382 2007

  11. [11]

    B., Abdo, A

    Atwood, W. B., Abdo, A. A., Ackermann, M., et al. 2009, ApJ, 697, 1071, doi: 10.1088/0004-637X/697/2/1071

    work page doi:10.1088/0004-637x/697/2/1071 2009

  12. [12]

    2024, MNRAS, 528, 976, doi: 10.1093/mnras/stae028

    Bhatta, G., Gharat, S., Borthakur, A., & Kumar, A. 2024, MNRAS, 528, 976, doi: 10.1093/mnras/stae028

    work page doi:10.1093/mnras/stae028 2024

  13. [13]

    2012, A&A, 548, A123, doi: 10.1051/0004-6361/201220056

    Biteau, J., & Giebels, B. 2012, A&A, 548, A123, doi: 10.1051/0004-6361/201220056

    work page doi:10.1051/0004-6361/201220056 2012

  14. [14]

    D., & Rees, M

    Blandford, R. D., & Rees, M. J. 1978, PhyS, 17, 265, doi: 10.1088/0031-8949/17/3/020

    work page doi:10.1088/0031-8949/17/3/020 1978

  15. [15]

    2015, MNRAS, 453, 1669, doi: 10.1093/mnras/stv1723

    Blinov, D., Pavlidou, V., Papadakis, I., et al. 2015, MNRAS, 453, 1669, doi: 10.1093/mnras/stv1723

    work page doi:10.1093/mnras/stv1723 2015

  16. [16]

    D., & Marscher, A

    Bloom, S. D., & Marscher, A. P. 1996, ApJ, 461, 657, doi: 10.1086/177092 B¨ ottcher, M., Reimer, A., Sweeney, K., & Prakash, A. 2013, ApJ, 768, 54, doi: 10.1088/0004-637X/768/1/54

    work page doi:10.1086/177092 1996

  17. [17]

    1996, Machine Learning, 24, 123, doi: 10.1023/A:1018054314350 —

    Breiman, L. 1996, Machine Learning, 24, 123, doi: 10.1023/A:1018054314350 —. 2001, Machine Learning, 45, 5, doi: 10.1023/A:1010933404324

    work page doi:10.1023/a:1018054314350 1996

  18. [18]

    2016, MNRAS, 462, 3180, doi: 10.1093/mnras/stw1830 de Prado, M

    Chiaro, G., Salvetti, D., La Mura, G., et al. 2016, MNRAS, 462, 3180, doi: 10.1093/mnras/stw1830 de Prado, M. L. 2018, Advances in Financial Machine Learning (Wiley)

    work page doi:10.1093/mnras/stw1830 2016

  19. [19]

    D., & Schlickeiser, R

    Dermer, C. D., & Schlickeiser, R. 1993, ApJ, 416, 458, doi: 10.1086/173251

    work page doi:10.1086/173251 1993

  20. [20]

    J., Pounds, K., et al

    Edelson, R., Turner, T. J., Pounds, K., et al. 2002, ApJ, 568, 610, doi: 10.1086/323779

    work page doi:10.1086/323779 2002

  21. [21]

    1993, ApJ, 407, 65, doi: 10.1086/172493

    Ghisellini, G., Padovani, P., Celotti, A., & Maraschi, L. 1993, ApJ, 407, 65, doi: 10.1086/172493

    work page doi:10.1086/172493 1993

  22. [22]

    2009, A&A, 503, 797, doi: 10.1051/0004-6361/200912303

    Giebels, B., & Degrange, B. 2009, A&A, 503, 797, doi: 10.1051/0004-6361/200912303

    work page doi:10.1051/0004-6361/200912303 2009

  23. [23]

    2013, Permutation Tests: A Practical Guide to Resampling Methods for Testing Hypotheses (New York: Springer)

    Good, P. 2013, Permutation Tests: A Practical Guide to Resampling Methods for Testing Hypotheses (New York: Springer)

    work page 2013

  24. [24]

    Nature , author =

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

    work page doi:10.1038/s41586-020-2649-2 2020

  25. [25]

    M., et al

    Hayashida, M., Nalewajko, K., Madejski, G. M., et al. 2015, ApJ, 807, 79, doi: 10.1088/0004-637X/807/1/79

    work page doi:10.1088/0004-637x/807/1/79 2015

  26. [26]

    Hunter, J. D. 2007, Computing in Science and Engineering, 9, 90, doi: 10.1109/MCSE.2007.55

    work page doi:10.1109/mcse.2007.55 2007

  27. [27]

    J., & Athanasopoulos, G

    Hyndman, R. J., & Athanasopoulos, G. 2018, Forecasting: Principles and Practice, 2nd edn. (Melbourne: OTexts) Katarzy´ nski, K., Ghisellini, G., Tavecchio, F., et al. 2005, A&A, 433, 479, doi: 10.1051/0004-6361:20041556

    work page doi:10.1051/0004-6361:20041556 2018

  28. [28]

    G., et al

    Kiehlmann, S., Savolainen, T., Jorstad, S. G., et al. 2016, A&A, 590, A10, doi: 10.1051/0004-6361/201527725 Kovaˇ cevi´ c, M., Chiaro, G., Cutini, S., & Tosti, G. 2020, MNRAS, 493, 1926, doi: 10.1093/mnras/staa394 K”unsch, H. R. 1989, The Annals of Statistics, 17, 1217

    work page doi:10.1051/0004-6361/201527725 2016

  29. [29]

    2015, ApJ, 804, 111, doi: 10.1088/0004-637X/804/2/111

    Joshi, M. 2015, ApJ, 804, 111, doi: 10.1088/0004-637X/804/2/111

    work page doi:10.1088/0004-637x/804/2/111 2015

  30. [30]

    2025, MNRAS, 539, 2185, doi: 10.1093/mnras/staf620

    Malik, Z., Akbar, S., Shah, Z., et al. 2025, MNRAS, 539, 2185, doi: 10.1093/mnras/staf620

    work page doi:10.1093/mnras/staf620 2025

  31. [31]

    1993, A&A, 269, 67, doi: 10.48550/arXiv.astro-ph/9302006

    Mannheim, K. 1993, A&A, 269, 67, doi: 10.48550/arXiv.astro-ph/9302006

    work page doi:10.48550/arxiv.astro-ph/9302006 1993

  32. [32]

    , keywords =

    Maraschi, L., Ghisellini, G., & Celotti, A. 1992, ApJL, 397, L5, doi: 10.1086/186531

    work page doi:10.1086/186531 1992

  33. [33]

    Marscher, A. P. 2014, ApJ, 780, 87, doi: 10.1088/0004-637X/780/1/87 18

    work page doi:10.1088/0004-637x/780/1/87 2014

  34. [34]

    P ., & Gear, W

    Marscher, A. P., & Gear, W. K. 1985, ApJ, 298, 114, doi: 10.1086/163592

    work page doi:10.1086/163592 1985

  35. [35]

    P., Jorstad, S

    Marscher, A. P., Jorstad, S. G., Larionov, V. M., et al. 2010, ApJL, 710, L126, doi: 10.1088/2041-8205/710/2/L126 M¨ ucke, A., Protheroe, R. J., Engel, R., Rachen, J. P., &

    work page doi:10.1088/2041-8205/710/2/l126 2010

  36. [36]

    2003, Astroparticle Physics, 18, 593, doi: 10.1016/S0927-6505(02)00185-8

    Stanev, T. 2003, Astroparticle Physics, 18, 593, doi: 10.1016/S0927-6505(02)00185-8

    work page doi:10.1016/s0927-6505(02)00185-8 2003

  37. [37]

    2013, MNRAS, 430, 1324, doi: 10.1093/mnras/sts711

    Nalewajko, K. 2013, MNRAS, 430, 1324, doi: 10.1093/mnras/sts711

    work page doi:10.1093/mnras/sts711 2013

  38. [38]

    2017, The Astronomer’s Telegram, 9928, 1

    Ojha, R., & Carpen, B. 2017, The Astronomer’s Telegram, 9928, 1

    work page 2017

  39. [39]

    2013, The Astronomer’s Telegram, 4941, 1

    Ojha, R., Carpenter, B., & Dutka, M. 2013, The Astronomer’s Telegram, 4941, 1

    work page 2013

  40. [40]

    1995, ApJ, 444, 567, doi: 10.1086/175631

    Padovani, P., & Giommi, P. 1995, ApJ, 444, 567, doi: 10.1086/175631

    work page doi:10.1086/175631 1995

  41. [41]

    Scikit-learn: Machine Learning in Python

    Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825, doi: 10.48550/arXiv.1201.0490

    work page Pith review doi:10.48550/arxiv.1201.0490 2011

  42. [42]

    Phipson, B., & Smyth, G. K. 2010, Statistical Applications in Genetics and Molecular Biology, 9, Article 39

    work page 2010

  43. [43]

    Platt, J. C. 1999, in Advances in Large Margin Classifiers, ed. A. J. Smola, P. Bartlett, B. Sch”olkopf, & D. Schuurmans (Cambridge, MA: MIT Press), 61–74

    work page 1999

  44. [44]

    G., Wilkinson, P

    Polatidis, A. G., Wilkinson, P. N., Xu, W., et al. 1995, ApJS, 98, 1, doi: 10.1086/192152

    work page doi:10.1086/192152 1995

  45. [45]

    N., & Romano, J

    Politis, D. N., & Romano, J. P. 1992, in Exploring the Limits of Bootstrap, ed. R. LePage & L. Billard (New York: Wiley), 263–270

    work page 1992

  46. [46]

    Newton and Charles J

    Politis, D. N., & Romano, J. P. 1994, Journal of the American Statistical Association, 89, 1303, doi: 10.1080/01621459.1994.10476870

    work page doi:10.1080/01621459.1994.10476870 1994

  47. [47]

    M., Villata, M., Acosta-Pulido, J

    Raiteri, C. M., Villata, M., Acosta-Pulido, J. A., et al. 2017, Nature, 552, 374, doi: 10.1038/nature24623

    work page doi:10.1038/nature24623 2017

  48. [48]

    2023, MNRAS, 519, 3000, doi: 10.1093/mnras/stac3701

    Sahakyan, N., Vardanyan, V., & Khachatryan, M. 2023, MNRAS, 519, 3000, doi: 10.1093/mnras/stac3701

    work page doi:10.1093/mnras/stac3701 2023

  49. [49]

    2015, PLoS ONE, 10, e0118432, doi: 10.1371/journal.pone.0118432

    Saito, T., & Rehmsmeier, M. 2015, PLoS ONE, 10, e0118432, doi: 10.1371/journal.pone.0118432

    work page doi:10.1371/journal.pone.0118432 2015

  50. [50]

    D., Norris, J

    Scargle, J. D., Norris, J. P., Jackson, B., & Chiang, J. 2013, ApJ, 764, 167, doi: 10.1088/0004-637X/764/2/167

    work page doi:10.1088/0004-637x/764/2/167 2013

  51. [51]

    2018, Research in Astronomy and Astrophysics, 18, 141, doi: 10.1088/1674-4527/18/11/141

    Shah, Z., Mankuzhiyil, N., Sinha, A., et al. 2018, Research in Astronomy and Astrophysics, 18, 141, doi: 10.1088/1674-4527/18/11/141

    work page doi:10.1088/1674-4527/18/11/141 2018

  52. [52]

    2017, MNRAS, 470, 3283, doi: 10.1093/mnras/stx1194

    Shah, Z., Sahayanathan, S., Mankuzhiyil, N., et al. 2017, MNRAS, 470, 3283, doi: 10.1093/mnras/stx1194

    work page doi:10.1093/mnras/stx1194 2017

  53. [53]

    A., Akbar, S., et al

    Shah, Z., Dar, A. A., Akbar, S., et al. 2025, Phys. Rev. D, 111, 123052, doi: 10.1103/61tz-jk8c

    work page doi:10.1103/61tz-jk8c 2025

  54. [54]

    C., & Rees, M

    Sikora, M., Begelman, M. C., & Rees, M. J. 1994, ApJ, 421, 153, doi: 10.1086/173633

    work page doi:10.1086/173633 1994

  55. [55]

    2015, MNRAS, 450, 183, doi: 10.1093/mnras/stv641

    Sironi, L., Petropoulou, M., & Giannios, D. 2015, MNRAS, 450, 183, doi: 10.1093/mnras/stv641

    work page doi:10.1093/mnras/stv641 2015

  56. [56]

    G., Staveley-Smith, L., de Blok, W

    Spada, M., Ghisellini, G., Lazzati, D., & Celotti, A. 2001, MNRAS, 325, 1559, doi: 10.1046/j.1365-8711.2001.04557.x

    work page doi:10.1046/j.1365-8711.2001.04557.x 2001

  57. [57]

    1991, ApJ, 374, 431, doi: 10.1086/170133

    Kuehr, H. 1991, ApJ, 374, 431, doi: 10.1086/170133

    work page doi:10.1086/170133 1991

  58. [58]

    K., & Yadav, K

    Tolamatti, A., Singh, K. K., & Yadav, K. K. 2023, MNRAS, 523, 5341, doi: 10.1093/mnras/stad1826

    work page doi:10.1093/mnras/stad1826 2023

  59. [59]

    M., & Padovani, P

    Urry, C. M., & Padovani, P. 1995, PASP, 107, 803, doi: 10.1086/133630

    work page internal anchor Pith review doi:10.1086/133630 1995

  60. [60]

    and Heckman, T

    Vaughan, S., Edelson, R., Warwick, R. S., & Uttley, P. 2003, MNRAS, 345, 1271, doi: 10.1046/j.1365-2966.2003.07042.x

    work page doi:10.1046/j.1365-2966.2003.07042.x 2003

  61. [61]

    2020 , note =

    Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Medicine, 17, 261, doi: 10.1038/s41592-019-0686-2

    work page doi:10.1038/s41592-019-0686-2 2020